tokfandomcom-20200215-history
Neural processing unit
An AI accelerator is a class of or computer system designed as for applications, especially s, and . Typical applications include algorithms for , and other -intensive or sensor-driven tasks. They are often designs and generally focus on arithmetic, novel s or capability. A number of vendor-specific terms exist for devices in this category, and it is an without a . AI accelerators can be found in many devices such as , tablets, and all around the world. History of AI acceleration s have frequently complemented the with special purpose accelerators for specialized tasks, known as s. Notable include s for s, s, s and . As and workloads rose in prominence in the , specialized hardware units were developed or adapted from existing products to these tasks. Early attempts As early as 1993, s were used as neural network accelerators e.g. to accelerate software. In the 1990s, there were also attempts to create parallel high-throughput systems for workstations aimed at various applications, including neural network simulations. -based accelerators were also first explored in the 1990s for both inference and training. ANNA was a neural net accelerator developed by . Heterogeneous computing refers to incorporating a number of specialized processors in a single system, or even a single chip, each optimized for a specific type of task. Architectures such as the have features significantly overlapping with AI accelerators including: support for packed low precision arithmetic, architecture, and prioritizing 'throughput' over latency. The Cell microprocessor was subsequently applied to a number of tasks including AI. In the , s also gained increasingly wide units, driven by video and gaming workloads; as well as support for low precision s. Use of GPU s or GPUs are specialized hardware for the manipulation of images and calculation of local image properties. The mathematical basis of neural networks and are similar, tasks involving matrices, leading GPUs to become increasingly used for machine learning tasks. , GPUs are popular for AI work, and they continue to evolve in a direction to facilitate deep learning, both for training and inference in devices such as s. GPU developers such as Nvidia are developing additional connective capability for the kind of dataflow workloads AI benefits from. As GPUs have been increasingly applied to AI acceleration, GPU manufacturers have incorporated hardware to further accelerate these tasks. Tensor are intended to speed up the training of neural networks. Use of FPGAs Deep learning frameworks are still evolving, making it hard to design custom hardware. devices such as s (FPGA) make it easier to evolve hardware, frameworks and software . Microsoft has used FPGA chips to accelerate . The application of FPGAs to AI acceleration motivated to acquire with the aim of integrating FPGAs in server CPUs, which would be capable of accelerating AI as well as tasks. Emergence of dedicated AI accelerator ASICs While GPUs and FPGAs perform far better than CPUs for AI related tasks, a factor of up to 10 in efficiency may be gained with a more specific design, via an (ASIC). These accelerators employ strategies such as optimized and the use of to accelerate calculation and increase of computation. Some adopted low-precision used AI acceleration are and the . In-memory computing architectures In June 2017, researchers announced an architecture in contrast to the based on and arrays applied to temporal detection, intending to generalize the approach to and systems. In October 2018, IBM researchers announced an architecture based on and to accelerate . The system is based on arrays. Nomenclature As of 2016, the field is still in flux and vendors are pushing their own marketing term for what amounts to an "AI accelerator", in the hope that their designs and s will become the . There is no consensus on the boundary between these devices, nor the exact form they will take; however several examples clearly aim to fill this new space, with a fair amount of overlap in capabilities. In the past when consumer s emerged, the industry eventually adopted s self-assigned term, "the GPU", as the collective noun for "graphics accelerators", which had taken many forms before settling on an overall pipeline implementing a model presented by . Potential applications * : Nvidia has targeted their boards at this space. * s * s, for example pesticide-free weed control. * , e.g. in mobile phones, a target for . * * s, e.g. navigation systems, e.g. the has been demonstrated successfully guiding autonomous drones. * s, increasing the range of tasks that can be automated, by adding adaptability to variable situations. * , to assist with diagnoses * , increasing the of s and ability to use increasingly advanced . * References Category:Neural processing unit